Active Learning-Assisted Human-In-The-Loop Framework for Rapid Borehole Image Annotation: Enabling AI-Powered Geological Interpretation at Scale
Abstract
Borehole image interpretation is a cornerstone of subsurface characterization in the oil and gas industry, providing high-resolution, azimuthally oriented data for in-situ stress estimation, wellbore stability assessment, and reservoir quality evaluation. Despite its geological significance, this process remains largely manual: expert geologists must individually review and annotate high-resolution image logs that can span thousands of feet per well, investing hundreds of hours per project before any automated prediction is available. This paper presents an integrated Human-in-the-Loop (HITL) active learning framework designed to overcome this annotation bottleneck. The system supports detection of geological features, particularly formation dips, from acoustic and resistivity image logs. It operates through three concurrent modules: (1) an interactive annotation interface overlaying model predictions on borehole images for expert review and correction; (2) an asynchronous background training engine leveraging the DINOv3 visual foundation model for deep feature extraction and continuous refinement of a lightweight convolutional segmentation head; and (3) an active learning selection module employing entropy-based uncertainty sampling that identifies and prioritizes the most informative depth intervals. Experimental results show that uncertainty sampling provides a measurable early-round advantage. Between 20 and 160 annotated images in the reference-well experiment, uncertainty sampling increases mean IoU by 1.74% and mean Dice/F1 by 1.44% relative to sequential depth-order selection. It also reaches a predefined early-convergence criterion of a mean angular error below 10° for two consecutive rounds, with a decrease in the confirming round, using 50 annotated images compared with 60 for sequential selection (16.7% fewer annotations). Validation across four wells, including the reference well, shows that uncertainty sampling reduces the annotation volume required to reach 95% of final IoU by 24.3% on average, with well-specific reductions ranging from 0% to 62.5%. In the architecture comparison, DINOv3 Projected Layers 3&6 outperforms DINOv3 Full Features across all three metrics, achieving 4.84° ± 0.25° angular error, 0.4409 ± 0.0071 IoU, and 0.6084 ± 0.0075 Dice/F1 with approximately 48,000 trainable parameters.